A method for diagnosing abnormal structural health monitoring data and an electronic device

Abstract

The present invention relates to the technical field of civil structure health monitoring, in particular to a method for diagnosing abnormal structural health monitoring data and an electronic device. The present invention transforms the problem of abnormal detection of structural health monitoring data into an image classification problem, that is, the acquired monitoring data is converted into multiple encoded images and then fused into a multi-channel image, realizing the whole process of automatic processing of abnormal diagnosis of multivariate data of structural health monitoring through CNN model training, time series encoded images, abnormal pattern learning. The whole process is intelligent, accurate, and the automation degree of the present invention is high. Except for the data marking process, the whole abnormal data diagnosis process is automatic processing, which greatly improves the processing speed of the present invention and can meet the real-time data preprocessing requirements of structural health monitoring online warning.

Description

Technical Field

[0001] The present invention belongs to the technical fields of convolutional neural networks, signal processing, and civil structural health monitoring, and particularly relates to a method for diagnosing anomalies in structural health monitoring data and an electronic device. Background Art

[0002] With the development of the economy, the demand for large-scale infrastructure is also increasing. The development and construction of long-span bridges, offshore platforms, super high-rise buildings, etc. are all expanding continuously. At present, most large-scale civil infrastructure facilities are equipped with comprehensive structural health monitoring (SHM) systems. The SHM system aims to ensure the safe and stable operation of large and complex engineering structures and provides a useful tool for the whole-life performance evaluation of structures. Among them, monitoring data is the basis of the entire SHM system, which can reflect the real-time behavior mechanism and performance evolution law of the structure. The reliability of the collected data has an important impact on the effectiveness of subsequent processing and evaluation. During the long-term service of these large-scale infrastructure structures, under the combined action of complex adverse factors such as harsh environmental interference, sensor failures, and monitoring system failures, it usually leads to the inevitable inclusion of various abnormal situations in the massive monitoring data. Abnormal monitoring data may lead to unreliable structural health assessments and safety warnings, and further lead to inappropriate maintenance decisions. Therefore, to overcome this problem, it is necessary to identify and eliminate anomalies from SHM data, which requires the data processing method to have high recognition accuracy and high computational efficiency. The online SHM system continuously collects a large amount of data, and there are generally various abnormal patterns in the data. If we manually detect abnormal data, it will consume a lot of time and effort, and even be an impossible task. Therefore, computer-aided data anomaly detection methods have emerged as the times require.

[0003] Data anomaly detection is a long-standing problem in SHM, and a large number of anomaly detection methods have been proposed to process the monitoring data collected from SHM systems. In these methods, the core algorithm is the calculation of anomaly indicators. Data anomalies are detected by comparing the indicators with their thresholds or sending the indicators into a pattern classifier. Usually, for each abnormal pattern, a specific index is required, and this index is defined based on expert experience, which makes the algorithm too cumbersome and lacks flexibility and robustness.

[0004] Since the existing traditional methods are difficult to handle the situation with multiple abnormal patterns, it is easy to produce the problems of under-processing and over-processing, unable to meet the accuracy and efficiency requirements of online early warning and structural state assessment, and the disadvantages of low automation degree and high cost of manual expert intervention. Therefore, there is a need for a method for diagnosing anomalies in structural health monitoring data and an electronic device based on convolutional neural networks and multi-channel coding maps that can intelligently and accurately monitor multiple abnormal patterns. Summary of the Invention

[0005] The object of the present invention is to overcome the above-mentioned deficiencies existing in the prior art, and to provide a method for diagnosing abnormal structural health monitoring data and an electronic device.

[0006] In order to achieve the above object of the invention, the present invention provides the following technical solutions:

[0007] A method for diagnosing abnormal structural health monitoring data includes the following steps:

[0008] S1: Obtain the monitoring data of the structure to be detected and convert it into a time-domain response image D{A};

[0009] S2: Randomly divide the time-domain response image D{A} into 3 data sets according to a preset ratio: the first training set D{A Tra}, the first validation set D{A V}, the first test set D{A Test}, and the calculation formula is as follows: D{A}=D{A Tra}+D{A V}+D{A Test};

[0010] S3: Manually annotate the 3 data sets according to the contour features of the time-domain image, select the data with manual annotation, and generate the second training set D{A Tra ,L}, the second validation set D{A V ,L} and the second test set D{A Test ,L};

[0011] S4: Encode the manually annotated monitoring data in the 3 data sets in step S3 into an encoded image group to obtain the third training set D E {A Tra ,L}, the third validation set D E {A V ,L} and the third test set D E {A Test ,L};

[0012] S5: Respectively fuse the encoded image groups in the 3 data sets in step S4 into corresponding multi-channel images to obtain the final training set D E-F {A Tra ,L}, the final validation set D E-F {A V ,L} and the final test set D E-F {A Test ,L};

[0013] S6: Build a CNN model, set the CNN model parameters, and use the final training set D E-F {ATra , L}, and the final validation set D E-F {A V , L} are input into the CNN model for model training; when the cross-entropy loss value of the CNN model converges or fits, stop training and output the current CNN model;

[0014] S7: Feed the final test set D E-F {A Test , L} into the CNN model, output the types and quantities of anomalies in the monitoring data in the final test set, and output a corresponding report. The present invention transforms the problem of anomaly detection in structural health monitoring data into an image classification problem, that is, the acquired monitoring data is converted into multiple encoded graphs and then fused into a multi-channel image, realizing the whole process of automatic processing of time series encoded graphs, CNN model training, anomaly pattern learning, and anomaly diagnosis of multivariate data in structural health monitoring. The whole process is intelligent, accurate, and the present invention has a high degree of automation. Except for a small part of the data marking process, the whole anomaly data diagnosis process is automatic processing, greatly improving the processing speed of the present invention and meeting the real-time data preprocessing requirements of structural health monitoring online warning.

[0015] As a preferred solution of the present invention, in step S1, the monitoring data is multivariate monitoring time series data in a preset time interval.

[0016] As a preferred solution of the present invention, the encoding operation in step S4 is to generate three encoded images: GASF (Gramian Angular Summation Field) image, MTF (Markov Transition Field) image, and URP (Unthresholded Recurrence Plots) image.

[0017] As a preferred solution of the present invention, the encoding of the GASF image includes the following steps:

[0018] S4.1.1: Obtain the time series X = {x1, x2,..., x i ,..., x n}, scale all its values within the interval [-1, 1] to obtain the scaled sequence where i ∈ [1, n], and n is the number of elements in the time series X;

[0019] S4.1.2: Represent the sequence in the polar coordinate system Encode the value of the sequence as the cosine of the angle and encode its timestamp as the radius through the following formula;

[0020]

[0021] wherein, is the cosine of the angle corresponding to element t, i is the timestamp of element N is a constant factor used to adjust the span of the polar coordinate system, is the set of natural numbers;

[0022] S4.1.3: Calculate the GASF image according to the following formula:

[0023]

[0024] wherein, I is a unit row vector, 'is the transpose operation of the corresponding matrix, and j ∈ [1, n].

[0025] As a preferred embodiment of the present invention, the encoding of the MTF image includes the following steps:

[0026] S4.2.1: Obtain the time series X = {x1, x2,..., x i ,..., x n}, determine the breakpoints {q1, q2,..., q Q} of Q equal-sized regions generated under the Gaussian curve, set Q quantile bins according to the breakpoints, and assign each element x i to the corresponding quantile bin q a ;

[0027] The breakpoints and the quantile bins follow the formula

[0028] wherein, P(q a+1 ) is the probability of the quantile bin q a+1 , P(q a ) is the probability of the quantile bin q a , Q is a positive integer less than n, i ∈ [1, n], n is the number of elements in the time series X, and a ∈ [1, Q];

[0029] S4.2.2: Calculate the transition between the quantile bins q j along the time axis through a first-order Markov chain method, and construct a Q×Q weighted adjacency matrix W; the expression of the weighted adjacency matrix W is:

[0030]

[0031] is the transition from the quantile bin q a to the quantile bin q bThe frequency values, where a ∈ [1, Q] and b ∈ [1, Q]; the diagonal represents the self-transition frequency value within the quantile bin.

[0032] S4.2.3: Normalize the weighted adjacency matrix W according to the formula ∑w i,j = 1 to obtain the Markov transformation matrix. Add the time position information of size N×N to the Markov transformation matrix, align each probability in chronological order, expand the Markov transformation matrix to obtain the Markov transition field M, and output it as the MTF image; the expression of the Markov transition field M is as follows:

[0033]

[0034] where M i,j|i-j＝k represents the probability that the point x at time interval k i is converted to x j , and the diagonal represents the self-transition probability of this point. 。

[0035] As a preferred solution of the present invention, the encoding of the URP image includes the following steps:

[0036] Obtain the time series X = {x1, x2,..., x i ,..., x n}}, and obtain the URP image according to the following formula:

[0037] URP i,j = ||x(i) - x(j)||;

[0038] where URP i,j is the URP image of the i-th element and the j-th element in the time series X, and ||·|| is the norm operation.

[0039] As a preferred solution of the present invention, when there are l encoded images in the multi-channel image in step S5, the fusion method is as follows:

[0040] Before fusion, keep the array sizes of the l images consistent, and then perform image fusion through weighted summation; the fusion formula is as follows:

[0041] I Fusion = α1·I1 + α2·I2 + α3·I3 +... + α l ·I l

[0042] where I Fusion is the fused image, I1, I2, I3,..., I l are different images, and α1, α2, α3,..., α lis the fusion weight value corresponding to the above image, and α1 + α2 + α3 +... + α l = 1.

[0043] As a preferred solution of the present invention, the CNN model includes an input layer, a convolutional layer, a batch normalization layer, a pooling layer, a flattening layer, a fully connected layer, a freezing layer, and a Softmax output layer connected in sequence.

[0044] As a preferred solution of the present invention, in step S6, the CNN model is trained in batches.

[0045] Among them, the batch size of the model training is 512, shuffle is enabled, the objective function is the cross-entropy function, the optimizer is the Adam optimizer, and the global maximum number of iterations, the initial learning rate α, and the decay learning rate are preset values.

[0046] An electronic device includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method described in any one of the above.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] The present invention transforms the problem of abnormal detection of structural health monitoring data into an image classification problem, that is, the acquired monitoring data is converted into multiple encoded images and then fused into a multi-channel image, realizing the whole process of automatic processing of time series encoded images, CNN model training, abnormal pattern learning, and abnormal diagnosis of multivariate data in structural health monitoring. The whole process is intelligent and accurate, and the automation degree of the present invention is high. Except in the data marking process, the whole abnormal data diagnosis process is automatic processing, greatly improving the processing speed of the present invention and meeting the real-time data preprocessing requirements of structural health monitoring online warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a schematic flowchart of a method for abnormal diagnosis of structural health monitoring data according to Embodiment 1 of the present invention;

[0050] Figure 2 is a framework flowchart of a practical application example of a method for abnormal diagnosis of structural health monitoring data according to Embodiment 3 of the present invention;

[0051] Figure 3 is an example of encoding a time series into an image and fusing a three-channel image in a method for abnormal diagnosis of structural health monitoring data according to Embodiment 3 of the present invention;

[0052] Figure 4 Schematic diagram of the detection result of structural health monitoring data anomaly diagnosis in the structural health monitoring data anomaly diagnosis method described in Embodiment 3 of the present invention;

[0053] Figure 5 Schematic diagram of the structure of an electronic device that utilizes the structural health monitoring data anomaly diagnosis method described in any one of Embodiments 1-3 according to Embodiment 4 of the present invention. Detailed implementation manners

[0054] The present invention will be further described in detail below in conjunction with test examples and specific implementation manners. However, it should not be understood that the scope of the above-mentioned subject matter of the present invention is limited to the following embodiments. All technologies implemented based on the content of the present invention belong to the scope of the present invention.

[0055] Embodiment 1

[0056] As Figure 1 shown, a structural health monitoring data anomaly diagnosis method includes the following steps:

[0057] S1: Obtain the monitoring data of the structure to be detected and convert it into a time-domain response image D{A}; the monitoring data is multivariate monitoring time series data in a preset time interval.

[0058] S2: Randomly divide the time-domain response image D{A} into 3 data sets according to a preset ratio: the first training set D{A Tra}, the first validation set D{A V}, and the first test set D{A Test}, and the calculation formula is as follows: D{A} = D{A Tra} + D{A V} + D{A Test}.

[0059] S3: Manually annotate the 3 data sets according to the contour features of the time-domain image, select the data with manual annotation, and generate the second training set D{A Tra ,L}, the second validation set D{A V ,L}, and the second test set D{A Test ,L}.

[0060] S4: Encode the manually annotated monitoring data in the 3 data sets in step S3 into an encoded image group to obtain the third training set D E {A Tra ,L}, the third validation set D E {A V ,L}, and the third test set D E {A Test ,L}.

[0061] S5: Respectively fuse the encoded image groups in the three data sets in step S4 into corresponding multi-channel images to obtain the final training set D E-F {A Tra ,L}, the final validation set D E-F {A V ,L} and the final test set D E-F {A Test ,L}.

[0062] When there are l types of encoded images, the fusion method of the multi-channel image is as follows:

[0063] Before fusion, keep the array sizes of the l types of images consistent, and then perform image fusion through weighted sum; the fusion formula is as follows:

[0064] I Fusion =α1·I1 + α2·I2 + α3·I3 +... + α l ·I l

[0065] where, I Fusion is the fused image, I1, I2, I3,..., I l are different images, α1, α2, α3,..., α l are the fusion weight values corresponding to the above images, and α1 + α2 + α3 +... + α l = 1.

[0066] S6: Build a CNN model, set the CNN model parameters, and input the final training set D E-F {A Tra ,L} and the final validation set D E-F {A V ,L} into the CNN model for model training; when the cross-entropy loss value of the CNN model converges or fits, stop training and output the current CNN model.

[0067] The CNN model includes an input layer, a convolutional layer, a batch normalization layer, a pooling layer, a flattening layer, a fully connected layer, a freezing layer, and a Softmax output layer connected in sequence. The CNN model uses batch processing for model training;

[0068] where, the batch size of the model training is 512, shuffle is enabled, the objective function is the cross-entropy function, the optimizer is the Adam optimizer, and the global maximum number of iterations, the initial learning rate α, and the decay learning rate are preset values.

[0069] S7: Input the final test set D E-F {A Test,L} is fed into the CNN model to output the types and quantities of anomalies in the monitoring data of the final test set, and a corresponding report is output.

[0070] Embodiment 2

[0071] The difference between this embodiment and Embodiment 1 is that in step S4, the encoding operation is to generate three encoded images, namely Gramian Angular Summation Field (GASF) image, Markov Transition Field (MTF) image, and Unthresholded Recurrence Plots (URP) image.

[0072] The basic principles of the GASF image and the MTF image are based on the following reference: "Wang Z, Oates T. Imaging time-series to improve classification and imputation[C] / / Twenty-Fourth International Joint Conference on Artificial Intelligence, 2015". The basic principle of the URP image is based on the following reference: "Sipers A, Borm P, Peeters R. On the unique reconstruction of a signal from its unthresholded recurrence plot[J]. Physics Letters A. 2011, 375(24): 2309-2321".

[0073] The encoding of the GASF image includes the following steps:

[0074] S4.1.1: Obtain the time series X = {x1, x2,..., x i ,..., x n}, scale all its values within the interval [-1, 1] to obtain the scaled sequence where i ∈ [1, n], and n is the number of elements in the time series X;

[0075] S4.1.2: Represent the sequence in the polar coordinate system Encode the values of the sequence as cosine of the angle through the following formula, and encode its timestamp as the radius;

[0076]

[0077] Among them, is the cosine of the angle corresponding to the element , t i is the element timestamp, and N is a constant factor used to adjust the span of the polar coordinate system;

[0078] S4.1.3: Calculate the GASF image according to the following formula:

[0079]

[0080] Among them, I is the unit row vector, 'is the transpose operation of the corresponding matrix, and i, j ∈ [1, n].

[0081] The encoding of the MTF image includes the following steps:

[0082] S4.2.1: Obtain the time series X = {x1, x2,..., x i ,..., x n} which has the characteristic of Gaussian distribution after normalization. First, determine the breakpoints that divide the area under the Gaussian curve into Q equal-sized regions. These breakpoints are the sorted list of Q = q1, q2,..., q Q and follow

[0083] where P(q a+1 ) is the probability of the quantile bin q a+1 , P(q a ) is the probability of the quantile bin q a , Q is a positive integer less than n, i ∈ [1, n], n is the number of elements in the time series X, and a ∈ [1, Q]. Thus, Q quantile bins are set, and each element x i is assigned to the corresponding quantile bin q a ;

[0084] S4.2.2: Calculate the transitions between the quantile bins q j along the time axis through the first-order Markov chain method to construct a Q×Q weighted adjacency matrix W; the expression of the weighted adjacency matrix W is:

[0085]

[0086] is the frequency value of the transition from the quantile bin q a to the quantile bin q b , a ∈ [1, Q], b ∈ [1, Q]; the diagonal is the self-transition frequency value within the quantile bin.

[0087] S4.2.3: Through ∑w i,jAfter normalization to 1, a Markov transformation matrix is obtained. The Markov transformation matrix is not sensitive to the dependence on the time step t i and the Markov matrix is extended by aligning each probability in chronological order to obtain the Markov transition field M, which is output as an MTF image; the expression of the Markov transition field M is as follows:

[0088]

[0089] where M i,j|i-j＝k represents the probability that the point x i at time interval k is transformed into x j . The diagonal (i.e., k = 0) is the self-transition probability of this point. For a time series of length N, the size of its Markov transition field M is N×N.

[0090] The encoding of the URP image includes the following steps:

[0091] Obtain the time series X = {x1, x2,..., x i ,..., x n}, and obtain the URP image according to the following formula:

[0092] URP i,j = ||x(i) - x(j)||;

[0093] where URP i,j is the URP image of the i-th element and the j-th element in the time series X, and ||·|| is a norm operation (such as the Euclidean norm or the maximum norm).

[0094] Example 3

[0095] This example is a specific application example of the method described in Example 2. As shown in the framework flowchart of Figure 2 , it specifically includes the following steps:

[0096] a: Select the data D{A} of a large structure A for one month, and randomly select and divide D{A} into three parts according to a certain ratio, D{A} = D{A Tra}+ D{A V}+ D{A Test}.

[0097] Convert this data into a time-domain response image to mark the abnormal types and corresponding quantities of this data set. Obtain D{A Tra , L}, D{A V , L} and D{A Test , L}.

[0098] b: Encode the grouped and classified data set into GASF images, MTF images, and URP images. Obtain D E {A Tra ,L}, D E {A V ,L} and D E {A Test ,L}.

[0099] Fuse the three encoded images one-to-one into a three-channel image. As Figure 3 shown. Obtain the training set D E-F {A Tra ,L}, the validation set D E-F {A V ,L} and the test set D E-F {A Test ,L}. In this embodiment, to uniformly retain the features of each image, α1 = α2 = α3.

[0100] c: Build a CNN model and set the corresponding CNN parameters. The specific network architecture and parameters are shown in Table 1.

[0101] Table 1: CNN Network Architecture and Parameters

[0102]

[0103]

[0104] When training the convolutional neural network model, batch processing is adopted, the batch size BatchSize is 512, and shuffle is enabled. The objective function selects the cross-entropy function. The global maximum number of iterations is 300 times, the initial learning rate α is 0.0001, the Adam optimizer is selected, and the decaying learning rate is adopted, with a decay rate of 0.98 / 50 steps.

[0105] Stop training the CNN model after the cross-entropy loss value converges or fits to obtain the trained CNN model.

[0106] Finally, feed the test set D E-F {A Test ,L} into the CNN model to automatically detect the types and quantities of data anomalies and output the corresponding report. As Figure 4 shown, its detection result is the confusion matrix of the predicted label and the actual label, where the numbers "1" to "7" represent different anomaly patterns, the rightmost column of the matrix represents the recall rate of the corresponding pattern, the last row of the matrix represents the accuracy rate of the corresponding pattern, and the lower right corner of the matrix represents the overall average recall rate and accuracy rate. According to Figure 4The results shown indicate that the global accuracy of the diagnosis results reaches 96.9%, the recall rate reaches 96.0%, and the accuracy rate reaches 96.42%. According to the above experimental results, it can be known that the present invention can simultaneously diagnose multi-source data for structural health monitoring with multiple abnormal patterns, and solves the multi-classification problem of multi-source data for structural health monitoring.

[0107] Embodiment 4

[0108] As Figure 5 shown, an electronic device includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, and when the instructions are executed by the at least one processor, the at least one processor is enabled to execute a method for diagnosing abnormalities in structural health monitoring data described in the foregoing embodiments. The input / output interface may include a display, a keyboard, a mouse, and a USB interface for inputting and outputting data; and a power supply is used to provide electrical energy for the electronic device.

[0109] Those skilled in the art can understand that all or part of the steps for implementing the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps including the above method embodiments; and the foregoing storage medium includes: removable storage devices, read-only memory (ROM), magnetic disks, or optical disks and other various media that can store program codes.

[0110] When the above integrated unit of the present invention is implemented in the form of a software functional unit and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the embodiments of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in various embodiments of the present invention. And the foregoing storage medium includes: removable storage devices, ROM, magnetic disks, or optical disks and other various media that can store program codes.

[0111] The foregoing is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for diagnosing abnormal structural health monitoring data, characterized in that It includes the following steps: S1: Obtain the monitoring data of the structure to be detected and convert it into a time-domain response image D{A}; S2: Randomly divide the time-domain response image D{A} into three datasets according to a preset ratio: the first training set D{A Tra}, the first validation set D{A V}, and the first test set D{A Test}, and the calculation formula is as follows: D{A} = D{A Tra}+D{A V}+D{A Test}; S3: Manually annotate the three data sets according to the contour features of the time-domain image, select the data with manual annotation, and generate the second training set D{A Tra , L}, the second validation set D{A V , L} and the second test set D{A Test ; S4: Encode the manually labeled monitoring data in the three datasets of step S3 into an encoded image group to obtain a third training set D E {A Tra , L}, a third validation set D E {A V , L} and a third test set D E {A Test , L}; S5: Fuse the encoded image groups in the three data sets in step S4 into corresponding multi-channel images respectively to obtain the final training set D E-F {A Tra , L}, the final validation set D E-F {A V , L} and the final test set D E-F {A Test , L}; S6: Build a CNN model, set the CNN model parameters, and input the final training set D E-F {A Tra , L} and the final validation set D E-F {A V , L} into the CNN model for model training; when the cross-entropy loss value of the CNN model converges or fits, stop training and output the current CNN model; S7: Feed the final test set D E-F {A Test , L} into the CNN model, output the types and quantities of anomalies in the monitoring data of the final test set, and output the corresponding report.

2. The abnormal diagnosis method for structural health monitoring data according to claim 1, characterized in that Wherein, The monitoring data in step S1 is multivariate monitoring time series data in a preset time interval.

3. The structural health monitoring data abnormal diagnosis method according to claim 2, characterized in that The encoding operation in step S4 is to generate three encoded images, namely the GASF image, the MTF image, and the URP image.

4. The structural health monitoring data anomaly diagnosis method according to claim 3, characterized in that, The encoding of the GASF image includes the following steps: S4.1.1: Obtain the time series X = {x1, x2,..., x i ,..., x n}, scale all its values within the interval [-1, 1] to obtain the scaled series where i ∈ [1, n], and n is the number of elements in the time series X; S4.1.2: Represent the sequence in the polar coordinate system Encode the value of the sequence as the cosine of an angle and its timestamp as the radius through the following formula; Among them, is the cosine of the angle corresponding to the element , t i is the timestamp of the element , N is a constant factor used to adjust the span of the polar coordinate system, is the set of natural numbers; S4.1.3: Calculate the GASF image according to the following formula: Where I is the unit row vector, 'is the transpose operation of the corresponding matrix, and j ∈ [1, n].

5. A method for abnormal diagnosis of structural health monitoring data according to claim 3, characterized in that, The encoding of the MTF image includes the following steps: S4.2.1: Obtain the time series X = {x1, x2,..., x i ,..., x n}, and determine the breakpoints {q1, q2,..., q Q} of Q equally sized regions generated under the Gaussian curve. Set Q quantile bins according to the breakpoints, and assign each element x i to the corresponding quantile bin q a ; The demarcation point and the quantile bin follow the formula where P(q a+1 ) is the probability of quantile bin q a+1 , P(q a ) is the probability of quantile bin q a , Q is a positive integer less than n, i ∈ [1, n], n is the number of elements in the time series X, and a ∈ [1, Q]; S4.2.2: Calculate the quantile bin q through a first-order Markov chain along the time axis j Between the conversions, construct a weighted adjacency matrix W of Q×Q; the expression of the weighted adjacency matrix W is as follows: is the quantile bin q a is converted to the quantile bin q b The frequency value, where a ∈ [1, Q] and b ∈ [1, Q]; the diagonal is the self-transition frequency value within the quantile bin; S4.2.3: Normalize the weighted adjacency matrix W according to the formula ∑w i，j = 1 to obtain the Markov transformation matrix. Add time position information of size N×N to the Markov transformation matrix, align each probability in chronological order, expand the Markov transformation matrix to obtain the Markov transition field M, and output it as the MTF image; the expression of the Markov transition field M is as follows: Among them, M i，j|i-j＝k represents the probability that the point x with a time interval of k i is converted to x j . The diagonal line represents the self-transition probability of this point.

6. The structural health monitoring data anomaly diagnosis method according to claim 3, characterized in that The encoding of the URP image includes the following steps: Obtain the time series X = {x1, x2,..., x i ,..., x n}, and obtain the URP image according to the following formula: URP i，j = ||x(i) - x(j)||; where URP i，j is the URP image of the \(i\)-th and \(j\)-th elements in the time series \(X\), \(\|\cdot\|\) is the norm operation, and \(i,j\in[1,n]\).

7. A method for abnormal diagnosis of structural health monitoring data according to claim 1, characterized in that, When there are l encoded images for the multi-channel image in step S5, its fusion method is: Keep the array sizes of the l images consistent before fusion, and then perform image fusion through weighted summation; the fusion formula is as follows: I Fusion = α1·I1 + α2·I2 + α3·I3 +... + α l ·I l ; Among them, I Fusion is the fused image, I1, I2, I3, ..., I l are different images, and α1, α2, α3, ..., α l are the fusion weight values corresponding to the above images, and α1 + α2 + α3 +... + α l = 1.

8. A method for abnormal diagnosis of structural health monitoring data according to claim 1, characterized in that, The CNN model includes an input layer, a convolutional layer, a batch normalization layer, a pooling layer, a flattening layer, a fully connected layer, a freezing layer, and a Softmax output layer connected in sequence.

9. The structural health monitoring data abnormal diagnosis method according to claim 1, characterized in that In step S6, the CNN model is trained using batch processing; Wherein, the batch size of the model training is 512, shuffle is enabled, the objective function is the cross-entropy function, the optimizer is the Adam optimizer, and the global maximum number of iterations, the initial learning rate α, and the decay learning rate are preset values.

10. An electronic device, characterized in that, It includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method according to any one of claims 1 to 9.

Images